Try Mistral Large 4: API Preview, Current Pricing, and Weight Plans

API preview now, weights later

Mistral announced Large 4's public preview October 6, 2026 with a Mistral Studio API entry. Developers can assess suitability now rather than waiting for weights to learn whether it fits.

This is not delivery of downloadable weights. Mistral promises them by month-end with further architecture and training details. As of October 7, the final download, license, and self-hosting instructions cannot be inferred from that plan.

The preview runs on Mistral's European infrastructure. Further global regions are future plans, not proof every region, account, or enterprise option is available.

Use the correct model ID

Enter Mistral Studio from the announcement and check model access, billing, and limits in your account. Public preview does not promise free credits to every account; console login is not unlimited calling eligibility.

The model page specifies mistral-large-4. Ordinary chat uses POST https://api.mistral.ai/v1/chat/completions, selecting that ID in model, conversation in messages, and a valid API key in the authentication header. See the official documents for complete requests.

For existing Mistral integrations, add Large 4 separately and retain the original fallback. A mistral-large-latest example does not establish a fixed Large 4 version. Confirm the actual selected model.

Reading the introductory price and estimating budgets

As of October 7, Standard Sale price is $0.68 input, $0.07 cached input, and $2.09 output per million tokens. Tokens measure model text usage; one token is not one Chinese character or document page.

The announcement card still shows $1.36 input and $4.18 output, while the pricing page lists these as original prices beside discounted rates. State that a current estimate uses sale pricing; do not mix original and sale rates in one calculation.

Meter input and output separately, applying cache prices only to eligible cached input. Repeated conversation history and multi-step agent tools can increase usage. Compare the total cost of completing the same task.

Standard, Batch, and Priority are separate tiers; these figures describe current Standard only. The October 6 changelog says the 50% launch discount lasts two weeks but gives no exact end time or zone. Do not invent a precise October 20 cutoff.

Estimate introductory and later regular pricing separately for long-term costs. Recheck tier and account billing before expansion instead of treating two-week pricing as permanent.

Where image/text and agent applications can benefit

Large 4 combines image understanding and text. Mistral emphasizes coding, complex documents, and agent workflows, and lists structured output, function calling, document questions, and Agents/Conversations interfaces.

Structured output returns agreed fields; function calling proposes tool requests. They help connect workflows but do not decide whether tools should have write, messaging, or payment permissions.

Hypothetical example: a seller extracts specifications from product descriptions and packaging photos. Use samples without customer data, checking fields, visual details, and missing information before integrating results into product records. This is not a measured result.

The 1M context listing permits long inputs, not guaranteed complete understanding of any document or free calls. Select relevant material instead of submitting everything at once to assess quality and cost more clearly.

Before switching, check three kinds of results

First, answer quality: compare representative samples for Chinese instructions, table fields, image details, and citations. Vendor benchmarks indicate positioning but cannot replace your business samples or promise universal improvements.

Second, application compatibility: inspect text, structured fields, tool calls, and streaming handling. Documentation says content can be text or typed content blocks; string-only applications need to account for this.

Third, task time and total cost: record input, output, and repeated calls before expanding. During preview, use separate configuration and retain the old path. For access/quota messages, check console and official guidance; changing egress does not establish resolved billing or eligibility.

What self-hosting teams can prepare now

Document the application's required inputs, outputs, and tool interfaces, using the preview to decide whether further evaluation is worthwhile. Separate evaluation from deployment plans to choose next steps when weights actually arrive.

Once weights, license, and deployment guidance exist, check hardware, inference frameworks, commercial conditions, and data requirements. The month-end promise still needs fulfillment; do not buy hardware on that basis or claim an ordinary VPS already runs the full model.